Professional Documents
Culture Documents
Dyn Part I
Dyn Part I
Fraydoun Rezakhanlou
Departmet of Mathematics, UC Berkeley
December 3, 2010
1
PART I
2
1 Systems of differential equations
Consider the differential equation
dx
(1.2) = f (x, t), x ∈ U, t ∈ R,
dt
where U ⊆ Rd is an open set, x : [t1 , t2 ] → U is a differentiable function and f : U × R → Rd
is a continuous function. The main question is this: Given an initial condition x(t0 ) = a,
does (2.1) possess a unique solution that is defined for all times? Before answering this, let
us examine some examples.
Example 1.1.
From the above examples we learn that some conditions are needed on the function f
in order to guarantee the existence of a unique globally defined solution for a given initial
condition. What is responsible for the blow-up in Example 1(i) is the fact that the velocity
or the growth rate is quadratic. Less is needed to have a blow-up as the following Exercise
indicates.
Exercise 1.2. Take any continuous function f : R → (0, ∞). Consider the equation
dx
dt
= f (x),
R ∞ subject to the initial condition x(0) = a. Show that we have a blow-up if and
dx
only if a f (x) < ∞.
p
The non-uniqueness in Example 1.1(ii) stems from the fact that the function f = |x|
has a cusp at 0, i.e., f 0 (0±) = ±∞. To avoid both cusp and super linear growth, it suffices
to assume that the function f is Lipschitz. More precisely, we say f is (uniformly) Lipschitz
if there exists a constant L such that for every x, y ∈ U and t ∈ R,
3
Theorem 1.3. Suppose f is Lipschitz. Then (1.1) has a unique solution for every initial
condition.
Lemma 1.4. Let x and y be two solutions with f satisfying (1.2). Then
Note that (1.3) implies the uniqueness part of Theorem 1.3 by assuming x(t0 ) = y(t0 ) = a.
d
(ϕ(t)e−2tL ) ≤ 0.
dt
This means that ϕ(t)e−2tL is a non-increasing function of t. Hence for t > t0 ,
This implies (1.3) for t > t0 . For t < t0 , first observe that we also have ϕ0 ≥ −2Lϕ. So, the
function ϕ(t)e2Lt is now non-decreasing. Hence,
Before turning to the question of existence, let us mention that in practice we may only
know an approximation of what appears on the right-hand side of (1.1). The following lemma
asserts that by solving (1.1) with an error both on the right-hand side and the initial data,
we are making a small error on the solution.
4
with |E1 | + |E2 | < . Then
L|t−t0 |
(1.4) |x(t) − y(t)| ≤ |x(t0 ) − y(t0 )|eL|t−t0 | + (e − 1).
L
We now turn to the question of existence. For simplicity let us assume that f (x, t) = f (x)
is independent of t. Also assume that x(0) = a. To find a solution, let us design an
approximation scheme. For n > 0, define xn (·) by the requirement that xn (0) = a, and
(
f xn nj if nj < t < j+1
n
dx n
, j ≥ 0,
(t) = j+1
j j+1
dt f xn n if n < t < n , j < 0.
Clearly such xn is piecewise linear and can be constructed. The existence of a solution is an
immediate consequence of Lemma 1.7.
Lemma 1.7. The sequence {xn } is Cauchy and if xn → x, then x solves (1.1).
Proof. First we establish the equicontinuity of the sequence {xn }. For this it suffices to show
j
is uniformly bounded for nj ≤ T . To see this, observe
j
that the sequence xn = xn n
that if j > 0, then
The case j < 0 can be treated likewise. By Ascoli’s theorem, we can find a convergent
subsequence of {xn }. We continue to write {xn } for such a subsequence. Note that
dxn
dt
− f (xn (t)) = |f (xn (j/n)) − f (xn (t))| ≤ L|xn (j/n) − xn (t)|
≤ Ln |f (xn (j/n))|,
j j+1
provided that t ∈ n
, n , j ≥ 0. Hence
Z t
1
xn (t) = a + f (xn (s))ds + O .
0 n
6
If xn → x, then Z t
x(t) = a + f (x(s))ds,
0
for t > 0. The case t < 0 can be treated likewise.
Exercise 1.8.
• (i) Carry out the proof for the time-dependent case.
• (ii) Given a square matrix A, define its norm kAk = maxv |Av|/|v|. Show that kA +
Bk ≤ kAk + kBk and kABk ≤ kAk kBk. Moreover, if A is a symmetric matrix, then
kAk is the absolute value of the largest eigenvalue of A.
Assuming that f is continuous and x-Lipschitz, we have showed the existence of a unique
solution. Let us write φtt0 (a) for such a solution, to display its dependence on the initial data
a. Since both x(t) = φtt0 (a) and y(t) = φtt1 (φtt10 (a)) solve (1.1) and satisfy x(t1 ) = y(t1 ), we
deduce
whenever t0 < t1 < t. This is the group property of the family {φtt10 }.
Remark 1.9. When U = Rd and f is Lipschitz, our existence proof implies that the
solutions exist for all time. This may not be true if U 6= Rd . If (t1 , t2 ) is the largest existence
interval with, say, t2 < ∞, then limt→t2 x(t) exists and belongs to ∂U .
Note that φtt0 = φ0t−t0 when f is independent of t. In this case, we simply write φt for φt0
Now (1.8) becomes
(1.9) φt ◦ φs = φt+s .
This certainly implies the Lipschitzness of φ in the x-variable. We can say more if f is a
differentiable function.
Note that the function (A, t) 7→ B(t)A satisfies the Lipschitz property (1.2) so long as t
is restricted to a bounded interval. Hence (1.12) must have a unique solution. Hence we
already have a candidate for Dφt (x), namely the solution of (1.12). From this we expect to
have the differentiability of φt with respect to x.
To turn the above heuristic reasoning into a rigorous proof, we go back to Lemma 1.7
and use xn (·). In fact the approximation scheme used in Lemma 1.7 produces a flow that is
denoted by φtn (x). One can readily check that Dφtn (x) exists. Indeed
j j
j j+1
d Df φ n (x) Dφn (x)
n n
t ∈ n
, n
, j ≥ 0,
(1.13) Dφtn (x) = j+1 j+1
dt Df φnn (x) Dφnn (x) t ∈ j , j+1 , j < 0.
n n
j
Note that by Lemma 1.7, if n
≤ T and |x| ≤ T , then
!
j
φnn (x) ≤ sup |f (x)| (eLT − 1) + T =: BT .
|x|≤T
We also set
CT = max{kDf (x)k : |x| ≤ BT }.
This implies that
j j j−1 j−1
Dφnn (x) ≤ Dφnn (x) − Dφnn (x) + Dφnn (x)
j−1 j−1
1
≤ C
n T
Dφnn (x) + Dφnn (x)
j−1
= 1 + n1 CT Dφnn (x)
j j
≤ · · · ≤ 1 + n1 CT ≤ e n CT ≤ eT CT .
8
From this and (1.13) we deduce the equicontinuity of Dx φn (x, t) in the t-variable. This
allows us to pass to the limit as n → ∞. Let ψ(x, t) be a limit along a subsequence for a
fixed x. Then
(
dψ
dt
(x, t) = Df (φt (x))ψ(x, t)
(1.14)
ψ(x, 0) = I.
But (1.14) has a unique solution. Hence the full sequence {Dx φ(n) } converges to ψ uniformly
in t for every x. On the other hand φ(n) converges uniformly to φ. Since the derivatives
converge to ψ, we deduce that Dx φ must exist and that Dx φ = ψ. As a result, φ is
differentiable in x.
As our next step we would like to show that indeed ψ = Dx φ is a continuous function.
Recall that we are assuming f ∈ C 1 . We have
Z t
kψ(x, t) − ψ(y, t)k ≤ (Df (φs (x))ψ(x, s) − Df (φs (y))ψ(y, s))ds + kψ(x, t0 ) − ψ(y, t0 )k.
t0
Then Z t Z t
τ (t) ≤ τ (t0 ) + C1 τ (s)ds + C2 kDf (φs (x)) − Df (φs (y))k.
t0 t0
Since Df is continuous, we learn that for every > 0, there exists δ > 0 such that if
|x − y| < δ, then
kDf (φs (x)) − Df (φs (y))k ≤
for s in a bounded interval. Hence for |x − y| < δ,
Z t
τ (t) ≤ τ (t0 ) + C1 τ (s)ds + C2 (t − t0 )
t0
for t > t0 and both t and t0 in a bounded interval. Using Grownwall’s inequality,
C2 C1 |t−t0 |
τ (t) ≤ eC1 |t−t0 | τ (t0 ) + (e − 1).
C1
9
Note τ (0) = 0. Thus
C2 C1 |t|
kψ(x, t) − ψ(y, t)k ≤ (e − 1)
C1
for |x − y| < δ. This proves the continuity of ψ.
So far we have established the theorem for the case k = 1. For higher k, consider the
system
dx dξ
(1.15) = f (x), = Df (x)ξ.
dt dt
If f is C 2 , then (f, Df ) is C 1 . Hence (1.15) has a C 1 -flow. If its flow is denoted by φ̂t , then
φ̂t (a, I) = (φt (a), Dφt (a)). Since ξ = Dx φt , we learn that φt is C 2 in x-variable. Using (1.15)
again we can show that φt is C 2 in both variables. This proves the theorem for k = 2. The
larger k can be treated by induction.
The equation (1.14) plays an important role in studying the stability of the equation
dx/dt = f (x). More precisely, if x, y are two solutions to
dx
(1.16) = f (x)
dt
d
with y = x + δv then dt
(δv) ≈ Df (x)(δv). Hence the relevant problem to study is now
dv
(1.17) = Df (x)v
dt
where x = x(t) is a solution to (1.16). This is closely related to (1.15) or (1.14) because
v(t) = Dφt (a)v(0) where a = x(0). If x(·) happens to be a fixed point a, then (1.17) becomes
dv
(1.18) = Av
dt
with A = Df (a). The equation (1.18) will be studied in Section 2. The study of the
equation (1.16) when x(·) is a periodic orbit, is the subject of Section 3 and is known as
Floquet Theory.
In the case of a discrete dynamical system, we start with a function f : Rd → Rd and set
xn+1 = f (xn ), so that xn = f n (a) for an initial choice a. Now if ξn = Df n (a), then
Imagine now that {yn } is another orbit and that yn is close to xn . If yn = xn + δvn , then
yn+1 = f (xn + δvn ) ≈ Df (xn )δvn + f (xn ). Since yn+1 = xn+1 + δvn+1 , we deduce that
10
δvn+1 = Df (xn )δvn . Motivated by this, let us study the non-autonomous linear dynamical
system
(1.20) vn+1 = Df (xn )vn .
provided that {xn } is an orbit of the original dynamical system xn+1 = f (xn ). If {xn } is
a fixed point, i.e., f (a) = a and xn = a for all n, then (1.20) becomes vn+1 = Avn for
A = Df (a). Hence vn = An v for an initial choice v. The behavior of vn as n → ∞ depends
on the eigenvalues of A and will be treated in Section 2. Again, if {xn } is a periodic orbit,
then the sequence {vn } is analyzed by Floquet Theory and will be discussed in Section 3.
2 Linear Systems
In this section we study linear dynamical systems. In the discrete case we have a d × d
matrix and we are interested in the behavior of the sequence (An x : n ∈ Z) for a nonzero
vector x ∈ Rd . In the continuous case we study the flow of the ODE
dx
(2.1) = Ax.
dt
We start with the discrete case. Using a Jordan normal form we can find a basis of Rd
such that the transformation x 7→ Ax has the following matrix representation:
A1 0
A2
(2.2) A=
. .
.
0 An
where each block is either of the form
λ 0 ... 0 0
1 λ . . . 0 0
(2.3) Aj = ..
.. ..
. . .
0 0 ... 1 λ
with λ a real eigenvalue of A, or else of the form
α −β 0 0 0
β α 0 0
1 0 α
−β
0 1 β α
...
1 0 α −β
0 0 1 β α
11
with λ = α + iβ, λ̄ = α − iβ a pair of complex eigenvalues of A. For each eigenvalue λ we
write Eλ for the corresponding invariant subspace. In the case of real λ,
Eλ = {v ∈ Rd : (A − λ)k v = 0 for some k ∈ N}.
Indeed if v ∈ Eλ , then (A − λ)k Av = A(A − λ)k v = 0. We now set
M M M
(2.4) E− = Eλ , E 0 = Eλ , E + = Eλ .
|λ|<1 |λ|=1 |λ|>1
Proof. By invariance, it suffices to verify the theorem when A is of the form (2.3) or
(2.2). First assume that A is of the form (2.2) and that x ∈ Eλ with |λ| < 1. For such a
transformation,
λx1
x1
.. x1 + λx2
T.= .. .
.
xd
xd−1 + λxd
To obtain a contraction, we would like to replace the off-diagonal entries with some small
number. To achieve this, let us try a diagonal change of coordinates of the form
x1 µ 1 x1
ϕ ... = ... .
xd µ d xd
We have
λx1
x1 µ1
.. µ2 x1 + λx2
−1
ϕ Tϕ . = .. .
.
xd µd−1
µd
xd−1 + λxd
This suggests choosing µi so that µ1
µ2
, . . . , µµd−1
d
are small. This can be done if µj = δ −j for a
small δ. Set T̂ = ϕ−1 T ϕ. It suffices to verify the theorem for
λx1
x1
.. δx1 + λx2
T̂ . = .. .
.
xd
δxd−1 + λxd
12
Here λ is real and |λ| < 1. Pick γ ∈ (0, |λ|). Then we can choose δ sufficiently small so that
|T̂ (x)| ≤ γ|x|,
simply because i (δxj + λxj+1 )2 ≤ δ 2 |x|2 + λ2 |x|2 + 2λδ|x|2 . Hence |T̂ n (x)| ≤ γ n |x| for some
P
0 < γ < 1.
The case |λ| > 1 can be treated likewise. Also for A of the form (2.4), a similar argument
applies.
Remark 2.1. Our proof indicates that the following is true: Pick γ such that γ < |λ|
for everyPeigenvalue P λ with |λ| < 1. Then there exists a basis {u1 , . . . , ur } for E + such
that if | i xi ui | = ( i x2i )1/2 denotes the standard norm with respect to this basis, then
|An x| ≤ γ n |x| for every n ∈ N and x ∈ E − . A similar comment applies to the space E + .
Exercise 2.2. Suppose A is of the form (2.3) with λ = ±1. Show that |An x| = O(nk−1 )
where k is the size of the matrix A. If A is of the form (2.4) with |λ| = 1, then show that
|An x| = O(nk−1 ) where now A is of the size 2k × 2k.
x
The situation is drastically different if T is not hyperbolic. For example if T 1 =
x2
−x1
, then the only eigenvalue is −1 and every orbit (T n (x) : n ∈ Z) with x 6= 0 is periodic
−x2
of period 2. To have another example, assume that T corresponds to a matrix of the form
R1 0
...
0 Rn
13
cos θi − sin θi
where each Ri = . In this case all eigenvalues are of norm 1 and if x =
sin θi cos θi
[x1 , . . . , xn ]t , with xj ∈ R2 , then T (x) = [R1 x1 , . . . , Rn xn ]t . Since each Ri is a rotation, the
torus
T(ρ1 , . . . , ρn ) = {(x1 , . . . , xn ) : |x1 | = ρ1 , . . . , |xn | = ρn },
is invariant. If we assume that all ρj ’s are nonzero, then T is an n-dimensional torus. The
effect of T on T(ρ1 , . . . , ρn ) is simply a translation in the following sense: If xj = ρj eiaj then
ρ1 ei(a1 +θ1 )
T (x) =
..
.
ρn ei(an +θn ) .
So in terms of angles, we simply have
(a1 , a2 , . . . , an ) 7→ (a1 + θ1 , . . . , an + θn ).
As we will see later, the orbits of this transformation are dense if θ1 , . . . , θn , 1 are rationally
independent. That is, if c1 θ1 + · · · + cn θn + cn+1 = 0 for integers c1 , . . . , cn+1 , then c1 = · · · =
cn+1 = 0.
We now turn to (2.1). We again use the form (2.2) with Aj as in (2.3) and (2.4). But
this time
M M M
(2.5) F+ = Eλ F 0 = Eλ F − = Eλ .
Re λ>0 Re λ=0 Re λ<0
Theorem 2.3. If x ∈ F + , then limt→+∞ |φt (x)| = +∞ and limt→−∞ |φt (x)| = −∞.
14
Let φt and φ̃t be the flow of A and à respectively. To compare φt with φ̃t , we calculate
d t −t 2
|φ φ̃ a| = 2(φt φ̃−t a) · (A − Ã)(φt φ̃−t a).
dt
Since
0
x1
..
x1
(A − Ã) . = −δ ,
..
.
xd
xd−1
we have
|(A − Ã)(x)| ≤ δ|x|.
As a result,
d t −t 2
|φ φ̃ a| ≤ 2δ|φt φ̃−t a|2 .
dt
Hence
|φt φ̃−t a| ≤ eδt |a|, t > 0,
for every a. This can be rephrased as
for t > 0. Now if Re λ < 0,then we have that limt→+∞ |φt a| = 0. Since
we deduce that
lim |φt a| = +∞,
t→+∞
whenever Re λ > 0.
The linear systems of the form we have studied so far can be used to study nonlinear
systems near a fixed point. Theorem 2.1 can be used to study the orbits of the system
xn+1 = f (xn ) near a fixed point. A celebrated theorem of Hartman–Grobman asserts that
the nonlinear system and its linearization near a hyperbolic fixed point are equivalent:
15
Theorem 2.4. Assume f is smooth with f (a) = a and A = Df (a) hyperbolic. Then there
exists a homeomorphism h from a neighborhood U of a onto a neighborhood of the origin
such that f = h−1 ◦ T ◦ h in U where T (x) = Ax.
Note that if vn = An v belongs to h(U ) then h−1 (An v) = xn belongs to U and {xn } is an
orbit for the f -system. The reason is simply f n = h−1 ◦ T n ◦ h.
Given the equation dx/dt = f (x) with f (a) = 0 and set A = Df (a). Let us write φt and
t
ψ for the flow associated with dx/dt = f (x) and dx/dt = Ax respectively. In the continuous
case, the analogue of Theorem 2.4 is this:
Theorem 2.5. Assume f is smooth and f (a) = 0. Assume that A = Df (a) is hyperbolic in
the sense that A has no purely imaginary eigenvalue. Then there exists a homeomorphism h
from a neighborhood of a onto a neighborhood of the origin such that
(2.10) φt = h−1 ◦ ψ t ◦ h.
Exercise 2.6. Given a matrix A, use Jordan Normal Form Theorem to find a collec-
tion of numbers l1 < l2 < · · · < lk , positive integers n1 , n2 , . . . , nk and linear subspaces
G1 , G2 , . . . , Gk such that dim Gj = nj and that if v ∈ (G1 ⊕ G2 ⊕ · · · ⊕ Gj ) − (G1 ⊕ · · · ⊕ Gj−1 )
then
1
lim log |An v| = lj .
n→∞ n
Observe that if
R1 0
A=
..
.
0 Rn
16
with
0 −βj
Rj = ,
βj 0
then
x1 ψt1 x1
φt ... = ... ,
xn ψtn xn
2 j cos βj t − sin βj t
where x1 , . . . , xn ∈ R and ψt z = z. Again the torus (4.5) is invariant
sin βj t cos βj t
and the flow on this torus is simply
In other words, we have a free motion on this torus. It turns out that the flow restricted
to Td is dense if and only if (β1 , . . . , βn ) are linearly independent over rationals. Consider
T : Td → Td defined by T (a1 , . . . , an ) = (a1 +θ1 , . . . , an +θn ). We have the following theorem.
Theorem 2.8. Assume that X is locally compact, second countable, with no isolated point.
Suppose T : X → X is continuous. Then the following statements are equivalent:
(ii) For any pair of nonempty open sets U and V , there exists an integer N ≥ 0 such that
T −N (U ) ∩ V 6= ∅.
Proof. (i) ⇒ (ii). Take x ∈ X with {T n (x) : n ∈ Z+ } is dense. Then there are infinitely
many indices n1 , n2 ∈ Z+ such that T n1 (x) ∈ U and T n2 (x) ∈ V . We may assume n1 ≥ n2
so that T n2 (x) ∈ V ∩ T −N (U ) for N = n1 − n2 .
(ii) ⇒ (i). Let {U1 , U2 , . . . } be a countable open base for X. Find n1 such that T −n1 U2 ∩
U1 6= ∅. Find an open set V2 such that V̄2 ⊆ T −n1 U2 ∩ U1 and V2 6= ∅. We then pick n2 ∈ Z+
17
such that T −n2 U3 ∩ V2 6= ∅ and a nonempty open set V3 such that V̄3 ⊆ T −n2 U3 ∩ V2 etc. By
induction, we find a sequence of open sets U1 = V1 ⊇ V̄2 ⊇ V2 ⊇ V̄3 ⊇ . . . with Vj 6= ∅ and
a sequence of positive integers such that V̄j T⊆ Vj−1 ∩ T −nj−1 Uj . Without loss of generality,
we may assume V̄1 is compact. Let A = ∞ j=1 V̄j . Evidently A 6= ∅ and if x ∈ A then
T nj (x) ∈ Uj+1 for all j, and x ∈ U1 . Hence (T n (x) : n ∈ Z+ ) is dense.
(ii) ⇒ (iii). Suppose U 6= ∅ is open and invariant. We have T −1 (U ) ⊆ U which implies
−N
T (U ) ⊆ U for N ≥ 0. Take a nonempty open set V . For some N ≥ 0 we have ∅ = 6
−N
T (U ) ∩ V ⊆ U ∩ V . Since U ∩ V 6= ∅ for every nonempty open set S V , the set U is dense.
(iii) ⇒ (ii). Let U and V be two nonempty open sets and set Û = n≥0 T −n (U ). Clearly
Û is open and invariant. Hence Û is dense and Û ∩ V 6= ∅. This implies T −n (U ) ∩ V 6= ∅ for
some n ≥ 0.
If T : X → X is a homeomorphism, then we can talk about the full orbit O(x) = (T n (x) :
n ∈ Z). The proof of Theorem 2.8 implies this.
(ii) For every nonempty open sets U and V , there exists N ∈ Z such that T N (U ) ∩ V 6= ∅.
We now show that the denseness of the full orbit is equivalent to the topological transivity.
Proof. Suppose that (T n (x) : n ∈ Z) is dense for some x. Let ω(x) and α(x) be the set of
limit points of O+ (x) and O− (x) respectively. Note that both ω(x) and α(x) are closed sets.
By assumption α(x) ∪ ω(x) = X. Hence either x ∈ ω(x) or x ∈ α(x). In the former case,
O(x) ⊆ ω(x) and this implies that ω(x) = X. In the latter case X = O(x) ⊆ α(x). By the
previous remark, we deduce that T is topologically transitive.
18
Proof of Theorem 2.7. Suppose θ1 , . . . , θn , 1 are not rationally dependent. Since a
translation of a dense set is dense, it suffices to show that the corresponding transformation
T is topologically transitive. Using Corollary 2.9 and Lemma 2.10, it suffices to show that
if U is open with T (U ) = U , then either U = ∅ or U is dense. Set f = 11U . We may regard
f : Rd → R as a periodic function (here we lifted f to a transformation on Rd .) Since
T (U ) = U , we have f ◦ T = f . Write the Fourier expansion of f :
X
f (x) = a(n1 , . . . , nd ) exp(2πi(n1 x1 + · · · + nd xd )).
n1 ,...,nd ∈Zd
From f = f ◦ T we learn
X X
a(n) exp(2πix · n) = a(n) exp(2πix · n) exp(2πiθ · n).
n n
a(n) = a(n)e2πiθ·n
are two nonempty invariant open sets. From Corollary 2.11 (T n (x) : n ∈ Z) is not dense for
every x ∈ Td .
Exercise 2.12. Show that if v = (v1 , . . . , vd ) and v1 , . . . , vd are not rationally dependent,
then (φ̂t (x) : t ∈ R+ ) is dense for every x ∈ Td . Conversely, if v1 , . . . , vd are rationally
dependent, then (φ̂t (x) : t ∈ R) is never dense when d ≥ 2.
Exercise 2.14. Show that the decimal expansion of 2n may start with any finite combination
of digits. (Hint: use T : T1 → T1 defined by T (x) = x + α(mod 1) with α = log10 2.)
19
Example 2.15. A function f : [0, 1] → R with f (0) = 0, f 0 (x) > 1 for all x ∈ [0, 1], induces
an expanding map T : T1 → T1 by T (x) = f (x)(mod 1). Such a function T is an example
of a chaotic dynamical system and its orbit structure is significantly more complex than
translations. As an example, take f (x) = mx with m > 1 an integer. If we identify T1 with
the set of complex numbers z such that |z| = 1, then T corresponds to the transformation
z 7→ z m .
A = {x : x = ·a1 a2 . . . ak ∗ ∗ ∗ . . . }, B = {x : x = ·b1 b2 . . . bk ∗ ∗ ∗ . . . }
From this and Theorem 4.8, we can readily deduce that T is topologically transitive.
Theorem 2.16 implies that some orbits are periodic and there exists a dense orbit. Do we
have any other type of an orbit? We now claim that, for example, when m = 3, then there
exists a point x for which ω(x) is the Cantor set. To see this, set
20
In fact h is continuous and strictly increasing except for points of finite expansion. Let us
write T m for z 7→ z m . We can now see that in fact h ◦ T3 = T2 ◦ h in K. Since T2 is
topologically transitive, there exists x with {T2n (x) : n ∈ N} dense. Each T2n (x) cannot
be a dyadic rational because dyadic rationals have finite orbits. Set y = h−1 (x). Then
{T3n y : n ∈ N} is dense in K.
In fact what we have shown for Tm is stronger than topological transitivity. Namely, Tm
is topologically mixing in the following sense:
If U and V are two nonempty open sets, then there exists N = N (U, V ) such that
−n
T (U ) ∩ V 6= ∅ for n > N .
3 Floquet Theory
In this section we study the orbits of a linear system with periodic coefficients. This is the
subject of the Floquet Theory.
In the discrete case, we are interested in the behavior of the sequence (xn : n ∈ N) with
xn+1 = An xn for a given collection of invertible matrices An satisfying
(3.1), An+N = An
for every n. In other words we have a non-autonomous system with periodic coefficients. If
the period N = 1 then An is independent of n and we studied the corresponding problem in
Section 2. In the continuous setting, we are interested in the flow of the dynamical system
dx
(3.2) = A(t)x
dt
with A satisfying
for all t.
Recall that if xn = f n (a), then its variation satisfies vn+1 = Df (xn )vn . If xn is a periodic
orbit of period N then we have (3.1) for An = Df (xn ). Similarly if we start with a nonlinear
problem of the form dxdt
= f (x) and look at its variation dvdt
= Df (x(t))v then A(t) = Df (x(t))
satisfies (3.3) whenever x(·) is periodic of period T .
We start with the discrete problem
Set
(3.5) Rn = An An−1 . . . A1 .
21
We certainly have
(3.6) vn = Rn v0 , Rn+N = Rn RN .
Proposition 3.1. Assume that all the eigenvalues of RN belong to the set {z : |z| < 1}.
Then limn→∞ vn = 0 exponentially fast. In fact limn→∞ Rn = 0 exponentially fast.
Set c0 = max{1, kA1 k, . . . , kAN −1 . . . A1 k}. If all the eigenvalues of RN belong to {z : |z| <
k
1}, then RN → 0 as k → +∞ exponentially fast. This completes the proof.
In this context, we would like to have a result similar to Theorem 2.1. It turns out that
Exercise 2.6 has a generalization.
Theorem 3.2. There exists a collection of numbers l1 < l2 < · · · < lk , positive integers
n1 , n2 . . . nk and linear subspaces G1 , . . . , Gk , such that n1 + · · · + nk = d, and if v ∈ (G1 ⊕
· · · ⊕ Gj ) − (G1 ⊕ · · · ⊕ Gj−1 ), then
1
lim log |Rn v| = lj .
n→∞ n
k
Proof. If n = kN + r with r ∈ {0, 1, . . . , N − 1}, then Rn = Rr RN . Hence
k
|Rn v| ≤ c0 |RN v|
22
We now apply Exercise 2.6 to RN .
Remark 3.3. In fact l1 < · · · < lk are the numbers N log |λ1 |, . . . , N1 log |λn | written in
1
23
This formula yields a candidate for f (R) for every analytic function f . If f is simply a
polynomial, then we already know how to calculate f (R) and we now would like to show
that this calculation is consistent with (3.8). We only need to verifyPthis when f (z) = z n .
The verification in this case is left to the reader. Also if f is given by ∞ n
0 an z over a region
containing γ, then (3.8) is verified by approximating f by polynomials.
We are now ready to define log R for an invertible matrix R. Since R is invertible,
0∈/ Spect(R). Pick a half-line L such that L ∩ Spect(R) = ∅. Set Ω = C − L. Take a branch
of log z defined in Ω. Use this branch for f in (3.7) to define C = log R. More precisely,
1 1 (−1)d−1 d−1
(3.9) log R = (log λ)I + B − 2 B 2 + · · · + B .
λ λ λd−1
This is simply obtained by using the expansion of log and using the fact B d = 0:
∞ j
1 X 1
log(λI + B) = log λ + log I + B = log λ + B (−1)j−1 .
λ j=1
λ
Even when R is a real matrix in Lemma 3.4, there might not exist a real C such that
C
e = R. For a real log R we need additional conditions.
Lemma 3.5. Let R be a real invertible matrix. There exists a real matrix C with eC = R if
R has no negative real eigenvalue. Moreover we can always find a real Z such that eZ = R2 .
Since R has no negative eigenvalue, we may choose the standard branch of log. That is
Ω = C − {x : x ≤ 0} and log(ρeiθ ) = log ρ + iθ for θ ∈ (−π, π). Note that log z = log ρ − iθ =
24
log z̄. As a result
−1
Z
C̄ = (z̄I − R)−1 log z̄dz̄.
2πi γ
Recall γ winds around Spect(R) once. Since R is real, Spect(R) = Spect(R). Hence the
curve γ̄ winds once clockwise around Spect(R). As a result
Z
1
C̄ = (zI − R)−1 log zdz = C.
2πi −γ̄
A1 0
For the existence of Z, without loss of generality we assume that R =
... is
0 Ak
Z1 0
its Jordan Normal Form. We find Z =
. .. such that eZj = A2j . If Aj corresponds
0 Zk
to a pair of complex conjugate eigenvalues α ± iβ with β 6= 0, then we can find Cj such
λ 0
...
1
that eCj = Aj with Cj real. We then set Zj = 2Cj in this case. If Aj = .
..
0 1 λ
2
corresponds to a real eigenvalue, then Aj has no negative eigenvalue, so we can find Zj such
that eZj = A2j .
(3.10) Rn = Pn Z n .
Also there exist real matrices P̂n and Ẑ such that P̂n+2N = P̂n and
(3.11) Rn = P̂n Ẑ n .
1
Proof. For (3.10) we simply choose Z = N
log RN and set
Pn = Rn Z −n .
25
We then have
The statement (3.10) is often phrased as the existence of a periodic change of coordinates
x = Pn y that transforms the system vn = Rn v0 to the system wn = Z n v0 . Note that the
latter is linear with constant coefficients.
We now turn to the continuous problem (3.2)–(3.3). First observe that if we solve the
matrix equation
(
d
dt
R(t) = A(t)R(t)
(3.12)
R(0) = I
dv(t)
= A(t)v(t), v(0) = v0 .
dt
We now argue that if (3.3) holds, then
This follows from the uniqueness; both R1 (t) = R(t + T ) and R2 (t) = R(t)R(T ) solve
(
dX
dt
= AX
X(0) = R(T ).
Proposition 3.7. If all the eigenvalues of R(T ) belongs to {z : |z| < 1}, then limt→+∞ R(t) =
0 exponentially fast.
The proof of Proposition 3.7 is very similar to the proof of Proposition 3.1 and is omitted.
26
Before stating and proving the analogue of Theorem 3.2, let us observe that R(t) is always
invertible. In fact we have a candidate for B = R−1 :
d d
B(t) = −R(t)−1 R(t)R(t)−1
dt dt
= −R(t)−1 A(t)R(t)R(t)−1
= −B(t)A(t).
Hence if B solves (
d
dt
= −B(t)A(t)
B(t)
B(0) = I,
then
d
BR = −BAR + BAR = 0.
dt
Hence BR = I, i.e., R is invertible. Obviously R(t) = φt0 and R(t)−1 = φ0t .
Theorem 3.8. There exists a collection of numbers l1 < · · · < lk , positive integers n1 , . . . , nk
and linear subspaces H 1 , . . . , H k , such that n1 + · · · + nk = d, and if v ∈ H 1 ⊕ · · · ⊕ H j −
H 1 ⊕ · · · ⊕ H j−1 , then
1
lim log |R(t)v| = lj .
t→+∞ t
R(t) = R(r)R(T )k ,
R(T )k = R(r)−1 R(t).
Hence
|R(T )k r| ≤ c1 |R(t)r|
where c1 = maxr∈[0,T ] kR(r)−1 k. From this and (3.14) we deduce that
1 1 1
lim log |R(t)r| = lim |R(T )k r|.
t→∞ t T k→∞ R
We now apply Exercise 2.6 to the matrix R(T ).
27
Remark 3.9. If λ1 , . . . , λn are the eigenvalues of R(T ), then the set {l1 , . . . , lk } coincides
with the set T1 log |λ1 |, . . . , T1 log |λn | .
Theorem 3.10. There exists matrices P (t) and C such that R(t) = P (t)etC and P (t + T ) =
P (t). Moreover there exist real matrices P̂ (t) and Ĉ such that P̂ (t + 2T ) = P̂ (t) and
R(t) = P̂ (t)etĈ .
Proof. Since R(T ) is invertible, there exists a matrix C such that R(T ) = eT C . Set
P (t) = R(t)e−T C . We have
Since R(2T ) = R(T )2 , we can find a real matrix Ĉ such that R(2T ) = exp(2T Ĉ). Set
P̂ (t) = R(t)e−tĈ . Then
P̂ (t + 2T ) = R(t + 2T )e−2T Ĉ e−tĈ
= R(t)R(T )2 e−2T Ĉ e−tĈ
= P̂ (t).
The eigenvalues of R(T ) are the Floquet multipliers. In practice it is hard to calculate
them. The following lemma is useful in some cases.
Lemma 3.11. We have a solution x(t) = p(t)λt with p(t + T ) = p(t) if and only if λT is an
eigenvalue of R(T ).
Proof. Suppose x(t) = p(t)λt is a solution with p a T -periodic function. Recall x(t) =
P (t)etC x0 with P (·) periodic. We have
P (t + T )e(t+T )C x0 = λT P (t)etC x0 ,
P (t)etC (eT C − λT I)x0 = 0.
This implies that eT C − λT I is not invertible. (Recall that R(t) = P (t)etC is invertible.)
Hence λT is an eigenvalue of R(T ) = eT C .
Conversely if λT is an eigenvalue of eT C , then we may choose µ such that µ is an eigenvalue
of C and Cx0 = µx0 , λT = eT µ . We then set p(t) = P (t)x0 so that p(t + T ) = p(t) and if
x(t) = p(t)etµ then
P (t)etC x0 = P (t)etµ x0 = p(t)etµ = x(t).
So x(t) is a solution.
28
Rt
tr(A(s))ds
Lemma 3.12. If z(t) = det R(t), then z(t) = e 0 .
Because of this,
d
det(R(t)) = tr A(t) det R(t),
dt
or ż(t) = tr(A(t))z(t) with z(0) = 1. This implies the lemma.
Show that v(t) = (− cos t, sin t)et/2 is a solution to (3.2). Find the Floquet multipliers. (Hint:
Use Lemma 3.9 and (3.16).)
Show that x̄(t) = (sin t, − cos t) is a solution. Find the Floquet multipliers for the variation
equation
dv
= Df (x̄(t))v.
dt
29
As it turns out, in the very low dimensional cases, the orbit structure of (6.1) cannot be
too complex.
Exercise 4.1. Let f : R → R, f (a) = f (b) = 0, a < b and f > 0 on (a, b). Show that for
every x ∈ (a, b),
ω(x) = {b}, α(x) = {a}.
In this section we study the orbits of (4.1) when d = 2. A lot can be said in this case
because of Jordan Curve Theorem, a closed simple curve in the plane divides the plane into
two connected components.
Theorem 4.2 (Poincaré–Bendixon). Assume ω(x) 6= ∅ and is bounded with no fixed point.
Then ω(x) is a closed orbit.
Theorem 4.3. Suppose that {φt (x) : t ≥ 0} is a subset of a closed bounded set K with K
having only finitely many fixed points. Then either ω(x) is a fixed point, or ω(x) is a closed
orbit, or ω(x) contains a finite number of fixed points and a set of orbits γ1 , . . . , γl with ω(γj )
and α(γj ) a fixed point.
Lemma 4.4. Suppose O+ (x) ∩ ω(x) 6= ∅. Then either x is a fixed point or a closed orbit.
To prepare for the proof of Lemma 4.4, let us make some observations. Suppose we have
an ODE in Rd as in (4.1) and let γ : U → Rd be a parametrization of a surface of codimension
one in Rd . We assume that γ(0) = x0 and that f (x0 ) = is not tangent to Γ = γ(U ). Define
F : U × R → Rd , U ⊆ Rd−1
by F (y, t) = φt (γ(y))
30
We have
because φ0 (x) = x and Dφ0 (x) = Identity. Since f (x0 ) is not tangent to Γ, we have that
DF (0, 0) is invertible. Hence in a neighborhood of (0, 0), say U0 × (−δ, δ) we have that F is
an invertible differentiable map with a differentiable inverse.
We now consider a planar dynamical system with Γ a curve transverse to f , i.e., f (x) is
not tangent to Γ at every x ∈ Γ. Let γ be a parametrization of Γ with I an interval. We
also assume that we have a flow box (chart) about Γ.
Lemma 4.5. Assume that γ : I → R2 with γ(I) = Γ transverse to f as above. Let {yn } be
a distinct collection of points in Γ with φtn (y0 ) = yn , and 0 < t1 < t2 < . . . . Then {yn } is a
monotone sequence on Γ, i.e., yn = γ(sn ) with {sn } a monotone sequence in I.
Proof. It suffices to show that y1 is between y0 and y2 . Let β be a simple curve made up of
{φt (y0 ) : 0 ≤ t ≤ t1 } and the segment L of Γ between y0 and y1 . By Jordan Curve Theorem,
β divides R2 into a bounded region Ω1 and unbounded region Ω2 .
Now either φt (y1 ) enters Ω1 , i.e., φt (y1 ) ∈ Ω1 for t > 0 and small or φt (y1 ) enters Ω2 .
Let us assume the former. Set E1 = {y ∈ L : φt (y) enters Ω1 , for t > 0} and E2 = L − E1 .
Since we have a flow chart about Γ, it is not hard to see that both E1 and E2 are open in L.
Since L is connected and by assumption E1 6= ∅, we deduce that φt (y) enters Ω1 for every
y ∈ L. From this we deduce that in fact {φt (y1 ) : t > 0} ⊆ Ω1 because φt (y1 ) cannot exit Ω1
through L, and cannot exit through β − L by the uniqueness of our ODE.
The complement of L in Γ consists of two connected arcs Γ0 and Γ1 with y0 an endpoint
of Γ0 and y1 an endpoint of Γ1 . If we can show that in fact Γ0 ⊆ Ω2 , Γ1 ⊆ Ω1 , we are
done because y2 = φt (y0 ) = φt−t1 (y1 ) ∈ Γ1 which means that y1 is between y0 and y2 . It
is not hard to see Γ1 ⊆ Ω1 because near Γ we have a flow box and in the box Γ1 and
{φt (y1 ) : t > 0, t small} belong to the same connected component.
31
Proof of Lemma 4.4. Assume that O+ (x) ∩ ω(x) 6= ∅. Let a ∈ O+ (x) ∩ ω(x) and if a
is a fixed point, then we are done. If a is not a fixed point, erect a transverse L through a.
Since a ∈ ω(x) = ω(a), there exist tn → +∞ with φtn (a) = an ∈ L and an → a. If an = am
for some n 6= m, then a is a periodic point which implies that x is a periodic point and we
are done. If an ’s are distinct, then by Lemma 6.5 {an } is a monotone sequence. But this is
impossible because limn→∞ an = a0 .
Recall that an invariant set A is called minimal if it has no proper invariant subset. As
a corollary to Lemma 4.4 we have this:
Corollary 4.6. If A is minimal and compact, then A is either a fixed point or a periodic
orbit.
Note that if A is compact and invariant, then we can use Zorn’s lemma to deduce that
A has a nonempty compact subset that is minimal. (Here we are using the fact that the
intersection of a finite number of invariant sets is again invariant.)
Let us state an exercise regarding the connectedness of ω(x):
Exercise 4.7.
T
(i) Show that ω(x) = τ >0 {φt (x) : t ≥ τ }.
(ii) Use (i) to show that if O+ (x) is bounded then ω(x) is connected.
Proof of Theorem 4.2. Assume that ω(x) is bounded with no fixed point. Let γ0 be
a minimal subset of ω(x). Then γ0 must be a periodic orbit by Corollary 4.6. Let a ∈ γ0
32
and erect a transverse L through a. Since a ∈ ω(x), we can find tn → +∞ such that
φtn (x) = xn ∈ L and xn → a. By Lemma 6.5, the sequence xn is monotone. (Note that if
xn = xm for some tn 6= tm , then x is a periodic point and a = xn for all n. In this case we
are done already.) If γ1 is another minimal set and if γ1 intersects L as well at a point b,
then xn → b and a monotone xn cannot converge to two distinct points a 6= b. Hence there
exists a neighborhood Ba of a so that Ba ∩ ω(x) = Ba ∩ γ0 . This is true for every a ∈ γ0 . By
compactness of γ0 , we can find a neighborhood B of γ such that B ∩ ω(x) = B ∩ γ0 . Since
ω(x) is connected, we must have ω(x) = γ0 .
Proof of Theorem 4.3. Assume that ω(x) is neither a fixed point nor a periodic orbit.
In fact the proof of Theorem 4.2 reveals that if ω(x) contains a periodic orbit, then it must
be equal to it. Hence ω(x) does not contain any periodic orbit. Since ω(x) is connected, it
cannot consist of fixed points only. Let y ∈ ω(x) is not a fixed point. Evidently ω(y) ⊆ ω(x),
α(y) ⊆ ω(x). To complete the proof, it suffices to show that ω(y) consists of a single fixed
point and the same is true for α(y). We only verify the former. Indeed if z ∈ ω(y) and
z is not a fixed point, then we can erect a transverse L at z. We can repeat the proof of
Theorem 4.2 to deduce that ω(x) ∩ L = ω(y) ∩ L = {z} because a transverse can only have
one point of ω(x). Also O+ (y) must intersect L at some point, say at y0 . But O+ (y) ⊆ ω(x).
So y0 = z. As a result, O+ (y) ∩ ω(y) 3 z and by Lemma 4.4, y must be a periodic point.
This contradicts our assumption that ω(x) is not a periodic orbit. Hence ω(y) is a fixed
point, completing the proof.
33
Example 4.8. Consider the system
(
ẋ1 = −x2 + x1 (1 − r2 ),
ẋ2 = x1 + x2 (1 − r2 ).
1 θ(t)
lim log ρ(t) = α, lim =β
t→∞ t t→∞ t
where θ is the lifted angle. Note that α ± iβ are the eigenvalues of A and α measures the
exponential rate of increase in ρ and β measures the linear growth rate of θ.
λ 0 cos θ
Exercise 4.10. Show that if A = in (4.1) and x(t) = ρ(t) , then 1t log ρ(t) → λ
1 λ sin θ
and 1t θ(t) → 0 as t → ∞.
u⊥ · u̇ = A(t)u · u⊥ .
35
Since u̇ = θ̇u⊥ , we obtain
(4.5) θ̇ = A(t)u · u⊥ .
Note that this equation is a first order nonlinear equation in θ and does not depend on ρ.
Multiplying both sides of (4.4) by u yields
(4.6) ρ̇ = ρ(A(t)u · u)
or
Rt
A(t0 )u(θ(t0 ))·u(θ(t0 ))dt0
(4.7) ρ(t) = e 0 ρ(0).
We are interested in
Z t
1 1
(4.8) lim log ρ(t) = lim A(s)u(s) · u(s)ds =: ρ̄
t→∞ t t→∞ t 0
1
(4.9) limθ(t) =: θ̄
t
t→∞
cos θ(s)
where for simplicity we wrote u(s) for u(θ(s)) = . In fact we already know what ρ̄
sin θ(s)
is. Recall that by Floquet Theory,
(4.10) x(t) = P (t)etC x(0)
with P (t) periodic in t, and e2T C = R(2T ) where R(t) is the fundamental
1 solution.1 If λ1 and
λ2 are the Floquet multipliers, then ρ̄ exists and belongs to the set T log |λ1 |, T log |λ2 | .
Let µ1 and µ2 denote the eigenvalues of the real matrix C. Then T1 log |λj | = Re µj . When
A(t) ≡ A is independent of t, then µ1 , µ2 are simply the eigenvalues of A.
Theorem 4.11. If A is T -periodic, then the rotation number θ̄ exists and equals Im µj or
− Im µj .
then the result follows from Exercise 4.10 and the preceding discussion. In this case θ̄ = Im µ
where µ = α + iβ or λ. If C is not in Jordan Normal Form, we can find a matrix Q such
that C = QĈQ−1 with Ĉ in Jordan Form. So,
36
cos θ(t)
tĈ
We already know that if e v = ρ(t) , then 1t θ(t) → θ̄ as t → +∞. Hence we
sin θ(t)
only need to make sure that if
cos θ(t) cos θ̂(t)
Q = ρ̂(t) ,
sin θ(t) sin θ̂(t)
Qz
then we still have 1t θ̂(t) → ±θ̄. For this define f : T1 → T1 by f (z) = |Qz| where z is a vector
of length 1. Clearly f is a continuous function. Also f is a homeomorphism because Q is
Q−1 z
invertible. Indeed f −1 (z) = |Q −1 z| . As a result, f has a continuous lift F : R → R. Then
θ̂(t) = F (θ(t)).
Here we regard T1 as the interval [0, 2π] with 0 = 2π an F enjoys the property F (θ + 2π) =
F (θ) ± 2π. (Note that we either have deg F = 1 or deg F = −1 by Lemma 5.3.) It is not
hard to see that
F (θ)
lim = ± deg F = ±1.
θ→±∞ θ
θ(t) θ̂(t)
Since we know that t
→ θ̄ as t → +∞, we learn that t
→ ±θ̄ as t → +∞, we learn that
θ̂(t)
t
→ ±θ̄. Of course we get θ̄ if the matrix Q does not reverse the orientation.
We now turn to the general periodic
case.
We know (4.10) and by the previous argument,
cos θ(t)
if y(t) = etC x(0) and y(t) = ρ(t) , then limt→+∞ θ(t) t
= θ̄ exists. We only need to
sin θ(t)
make sure that the matrices P (t) do not change the angles by much, i.e., if
cos θ(t) cos θ̂(t)
P (t) = r(t) ,
sin θ(t) sin θ̂(t)
θ̂(t) θ(t)
lim = lim .
t→∞ t t→∞ t
37
Similarly if deg F ≡ −1, then
θ̂(t) θ(t)
lim = − lim .
t→∞ t t→∞ t
38